Intelligent simulation diagnosis method and system for power failure

Abstract

The invention relates to the technical field of power system simulation, and discloses a power failure intelligent simulation diagnosis method and system, and the method comprises the steps: collecting an operation data sequence, adjusting the parameters of a pre-established power system digital twin model through an adaptive parameter estimation algorithm, and generating a dynamic simulation model; sequentially injecting a plurality of preset fault types, running simulation to generate a full-time sequence simulation data sequence corresponding to each preset fault type, and extracting a plurality of time sequence feature vectors; respectively calculating the matching degree between the real-time operation data sequence and the full-time sequence simulation data sequence of each preset fault type on a plurality of time sequence feature vectors to obtain a plurality of matching degree indexes; fusing the plurality of matching degree indexes according to a preset weight to generate a comprehensive similarity evaluation value; and inputting the comprehensive similarity evaluation value into a fault classification model, and outputting a diagnosis result. According to the invention, the adaptability and precision of power fault diagnosis can be improved.

Description

Technical Field

[0001] This invention relates to the field of power system simulation technology, and in particular to an intelligent simulation diagnosis method and system for power faults. Background Technology

[0002] As the core infrastructure for energy supply, the stability of the power system directly affects the continuity of industrial production and the protection of people's livelihoods. With the increasing proportion of renewable energy connected to the grid and the growing complexity of the power grid topology, the probability of various faults such as short-circuit faults and equipment malfunctions has significantly increased. Failure to diagnose fault types accurately and promptly can easily trigger a chain reaction, leading to widespread power outages and causing huge economic losses. Therefore, achieving rapid and accurate diagnosis of power faults is a key technological requirement for ensuring the safe and reliable operation of the power system.

[0003] Traditional power fault diagnosis methods primarily rely on manual experience combined with local monitoring data, resulting in low diagnostic efficiency, high subjectivity, and a high misjudgment rate, making them ill-suited for the dynamic operation of complex power grids. In recent years, diagnostic technologies combining data-driven approaches and simulation modeling have emerged, but they still have several shortcomings: simulation models cannot match real-time changes in the operating status of the power system caused by load fluctuations, equipment aging, and other factors, leading to discrepancies between simulation and actual operating data and affecting diagnostic accuracy; fault characteristics lack the ability to fully explore the temporal correlations throughout the fault's lifecycle, making it difficult to comprehensively depict the fault evolution pattern. Therefore, it is necessary to design a technical solution that can overcome these deficiencies. Summary of the Invention

[0004] To address the issues of poor adaptability and insufficient accuracy in traditional power fault diagnosis, this invention proposes an intelligent simulation diagnosis method and system for power faults, which can improve the adaptability and accuracy of power fault diagnosis.

[0005] The technical solution adopted in this invention is as follows: A power fault intelligent simulation diagnosis method includes: Collect real-time operating data sequences of the power system within a preset time window, combine the real-time operating data sequences with an adaptive parameter estimation algorithm, adjust the parameters of the pre-established digital twin model of the power system, and generate a calibrated dynamic simulation model. In the dynamic simulation model, multiple preset fault types are injected sequentially, and the simulation is run to generate a full-time simulation data sequence corresponding to each preset fault type. Then, multiple time-series feature vectors are extracted from the full-time simulation data sequence. The matching degree between the real-time running data sequence and the full-time simulation data sequence for each preset fault type is calculated on multiple time-series feature vectors to obtain multiple matching degree indices; Multiple matching indicators are fused according to preset weights to generate a comprehensive similarity evaluation value. The comprehensive similarity evaluation value is then input into a pre-trained fault classification model to output the actual fault type diagnosis result of the power system.

[0006] Furthermore, the acquisition of real-time operating data sequences of the power system within a preset time window includes: Synchronously collect real-time operational data sequences from multiple monitoring points within a preset time window to form an original operational data set. The real-time operational data sequences include voltage data sequences, current data sequences, and frequency data sequences. Real-time quality assessment and filtering are performed on the original operational data set, including data integrity verification, validity verification, and consistency verification. For abnormal data points that fail the verification, they are filled and corrected based on historical data from the same monitoring point or real-time data from adjacent monitoring points to generate a real-time running data sequence.

[0007] Furthermore, the step of combining real-time running data sequences and adaptive parameter estimation algorithms to adjust the parameters of the pre-established digital twin model of the power system and generating a calibrated dynamic simulation model includes: The real-time running data sequence is input into the pre-established digital twin model of the power system, and the equivalent parameters of key components in the digital twin model of the power system are dynamically identified and adjusted based on the online recursive parameter identification algorithm. The key components include generators, lines and loads. Run the digital twin model of the power system using the adjusted parameters to generate the corresponding simulation output sequence; Calculate the normalized root mean square error between the simulation output sequence and the real-time running data sequence within the same time period. If the error is greater than the preset convergence threshold, repeat the above two steps for iterative calibration. When the error is less than or equal to the convergence threshold, the model under the current parameters is locked as the calibrated dynamic simulation model.

[0008] Furthermore, the method of dynamically identifying and adjusting the equivalent parameters of key components in the digital twin model of the power system based on the online recursive parameter identification algorithm includes: The recursive least squares method with a forgetting factor is used as the online recursive parameter identification algorithm; The real-time running data sequence is captured by a sliding time window of preset length, and the parameter identification input-output data pairs of key components are constructed. For each key component to be identified, the parameter variability of the key component is calculated based on the deviation between its historical operating data and simulation data; Based on the parameter variability, dynamically determine the subset of key components that need parameter updates within the current time window; Recursive parameter identification and adjustment are performed only on components in a subset of key components to update the equivalent parameters in the power system digital twin model.

[0009] Furthermore, the step of sequentially injecting multiple preset fault types into the dynamic simulation model and running the simulation to generate a full-time simulation data sequence corresponding to each preset fault type includes: On the calibrated dynamic simulation model, configure multiple preset fault types to be injected, and set the fault parameters, injection time and duration for each fault type; The simulation environment of the calibrated dynamic simulation model is initialized based on the running status of the real-time running data sequence at the end of the preset time window. With the set simulation step size, the injection and simulation of each preset fault type are executed sequentially, and the simulation data of the entire process from the pre-fault steady state, the fault occurrence, the fault duration to the fault clearance and recovery are recorded synchronously. The full-process simulation data corresponding to each preset fault type is integrated and time-aligned to generate a full-time simulation data sequence.

[0010] Furthermore, the extraction of multiple time-series feature vectors from the full-time simulation data sequence includes: The full-time simulation data sequence is segmented by a sliding window to obtain multiple simulation data subsequences; For each subsequence of simulation data, calculate multiple initial feature indices in the time domain, frequency domain, and time-frequency domain respectively; Arrange the same initial characteristic index of all simulation data subsequences under the same fault type in chronological order to form the time series change curve of the characteristic index; From all the time-series variation curves, select the curves with the highest correlation to typical failure modes, and use their corresponding characteristic indicators as the basic features. Based on the aforementioned basic features, dimensionality reduction and reconstruction are performed using principal component analysis to generate multiple time-series feature vectors corresponding to each preset fault type.

[0011] Furthermore, the step of calculating the matching degree between the real-time running data sequence and the full-time simulation data sequence for each preset fault type on multiple time-series feature vectors yields multiple matching degree indices, including: For each preset fault type, multiple time-series feature vectors corresponding to the real-time running data sequence and the full-time simulation data sequence are extracted respectively; For each time-series feature vector, a dynamic time warping algorithm is used to calculate the minimum warping distance between the feature vector time series of the real-time running data sequence and the corresponding feature vector time series of the full-time simulation data sequence. The minimum normalization distance is mapped to a value between 0 and 1 using a preset normalization function, and is used as the single matching degree under this time series feature vector; For each preset fault type, its single matching degree on all time-series feature vectors is summarized to form multiple matching degree indices for that fault type.

[0012] Furthermore, the step of fusing multiple matching indicators according to preset weights to generate a comprehensive similarity evaluation value includes: Obtain multiple matching metrics corresponding to each preset fault type; Construct a matching degree index matrix, where rows correspond to different preset fault types, columns correspond to different time-series feature vectors, and matrix elements are the corresponding matching degree indices; For each time-series feature vector, calculate its corresponding matching degree index, which has the discriminant and discreteness among all preset fault types; Based on the entropy weight method, the fusion weight of each time series feature vector is dynamically calculated according to the discrimination and dispersion, and used as the preset weight; Using the calculated fusion weights, the matching degree indices in each row of the matching degree index matrix are weighted and summed to generate a comprehensive similarity evaluation value for each preset fault type.

[0013] Furthermore, the step of inputting the comprehensive similarity evaluation value into the pre-trained fault classification model and outputting the actual fault type diagnosis result of the power system includes: Construct a pre-trained fault classification model based on a deep neural network, where the number of nodes in the input layer is the same as the dimension of the comprehensive similarity evaluation value, and the number of nodes in the output layer is the same as the number of preset fault types. Using simulation data from historical multi-fault scenarios, a comprehensive similarity evaluation value sample set for different fault types and their corresponding real fault type labels are generated to train and cross-validate the fault classification model. The comprehensive similarity evaluation value is input into the pre-trained fault classification model, and its softmax layer outputs the probability distribution corresponding to each preset fault type. The fault type with the highest probability value is selected as the candidate diagnostic result, and it is checked whether the probability value exceeds the preset confidence threshold. If it exceeds the threshold, the candidate diagnostic result is output as the actual fault type diagnostic result of the power system. If it does not exceed the threshold, an indication signal for an unknown or compound fault is output.

[0014] A power fault intelligent simulation diagnosis system, comprising: The dynamic simulation model building module is configured to collect real-time operating data sequences of the power system within a preset time window, combine the real-time operating data sequences with an adaptive parameter estimation algorithm, adjust the parameters of the pre-established digital twin model of the power system, and generate a calibrated dynamic simulation model. The time-series feature vector extraction module is configured to sequentially inject multiple preset fault types into the dynamic simulation model, run the simulation to generate a full-time simulation data sequence corresponding to each preset fault type, and then extract multiple time-series feature vectors from the full-time simulation data sequence. The matching degree index calculation module is configured to calculate the matching degree between the real-time running data sequence and the full-time simulation data sequence of each preset fault type on multiple time-series feature vectors, and obtain multiple matching degree indices. The fault type diagnosis result output module is configured to fuse multiple matching degree indicators according to preset weights to generate a comprehensive similarity evaluation value, and input the comprehensive similarity evaluation value into a pre-trained fault classification model to output the actual fault type diagnosis result of the power system.

[0015] The beneficial effects of this invention are as follows: This invention effectively solves the problems of poor adaptability and insufficient accuracy in traditional power fault diagnosis. By dynamically calibrating the digital twin model of the power system, it achieves accurate matching between the simulation model and the real-time operating state of the system, significantly reducing simulation deviations caused by fixed-parameter models and providing reliable data support for subsequent diagnosis. Combining feature mining and multi-dimensional matching degree evaluation of full-time simulation data, it comprehensively captures the evolution patterns of faults, overcoming the limitations of diagnosis based on single features or local data, and improving the ability to distinguish similar fault types. Employing weighted fusion similarity evaluation and a pre-trained classification model enhances the reliability and generalization of diagnostic results, enabling accurate identification of known fault types as well as effective alerts for unknown or complex faults. Attached Figure Description

[0016] Figure 1 This is one of the flowcharts of the intelligent simulation diagnosis method for power faults in Embodiment 1 of the present invention.

[0017] Figure 2 This is the second flowchart of the intelligent simulation diagnosis method for power faults in Embodiment 1 of the present invention. Detailed Implementation

[0018] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] Example 1 like Figure 1 As shown, this embodiment provides an intelligent simulation diagnosis method for power faults, including: Collect real-time operating data sequences of the power system within a preset time window, combine the real-time operating data sequences with an adaptive parameter estimation algorithm, adjust the parameters of the pre-established digital twin model of the power system, and generate a calibrated dynamic simulation model. In the dynamic simulation model, multiple preset fault types are injected sequentially, and the simulation is run to generate a full-time simulation data sequence corresponding to each preset fault type. Then, multiple time-series feature vectors are extracted from the full-time simulation data sequence. The matching degree between the real-time running data sequence and the full-time simulation data sequence for each preset fault type is calculated on multiple time-series feature vectors to obtain multiple matching degree indices; Multiple matching indicators are fused according to preset weights to generate a comprehensive similarity evaluation value. The comprehensive similarity evaluation value is then input into a pre-trained fault classification model to output the actual fault type diagnosis result of the power system.

[0020] like Figure 2 As shown, this intelligent simulation diagnosis method for power faults can be implemented using the following steps: S1: Collect the real-time operating data sequence of the power system within a preset time window. The real-time operating data sequence includes voltage data sequence, current data sequence and frequency data sequence. S2: Based on the real-time running data sequence, adjust the parameters of the pre-established digital twin model of the power system through an adaptive parameter estimation algorithm to generate a calibrated dynamic simulation model; S3: In the dynamic simulation model, multiple preset fault types are injected sequentially, and the simulation is run to generate a full-time simulation data sequence corresponding to each preset fault type; S4: Extract multiple time-series feature vectors from the full-time simulation data sequence; S5: Calculate the matching degree between the real-time running data sequence and the full-time simulation data sequence of each preset fault type on multiple time-series feature vectors to obtain multiple matching degree indices; S6: Merge multiple matching metrics according to preset weights to generate a comprehensive similarity assessment value; S7: Input the comprehensive similarity evaluation value into the pre-trained fault classification model and output the actual fault type diagnosis result of the power system.

[0021] For example, a real-time operating data sequence of the power system is collected within a preset time window. The preset time window refers to a fixed period of time during which the operating data of the power system is continuously collected. It can be selected as 5 minutes or 10 minutes. The real-time operating data sequence is a core data set reflecting the operating status of the system. It includes voltage data sequence, which refers to the continuous data of voltage changes over time at key nodes of the system; current data sequence, which refers to the continuous data of current changes over time in the main power supply branches; and frequency data sequence, which refers to the continuous data of the overall power supply frequency of the system fluctuating over time. The collection process can be completed by a monitoring device deployed at the core of the system. The monitoring device can be a high-precision power quality analyzer, which directly collects the above three types of basic operating data and integrates them into the original sequence.

[0022] Based on the real-time operational data sequence, the parameters of the pre-established power system digital twin model are adjusted using an adaptive parameter estimation algorithm, thereby generating a calibrated dynamic simulation model. The power system digital twin model is a model that digitally replicates the physical structure and operating characteristics of the power system, including the basic equivalent parameters of core components such as generators, lines, and loads. The adaptive parameter estimation algorithm is an algorithm that can dynamically adjust the model parameters based on real-time data. Its execution process involves inputting the real-time operational data sequence into the model, comparing the model output with the real-time data through the algorithm, and adjusting the equivalent parameters of key components in the model accordingly. For example, the transient reactance of the generator can be adjusted to 0.18 pu, and the resistance parameter of the line can be adjusted to 0.01 Ω / km, until the deviation between the model output and the real-time data meets the preset requirements.

[0023] In the calibrated dynamic simulation model, multiple preset fault types are injected sequentially and the simulation is run to generate a full-time simulation data sequence corresponding to each preset fault type. The preset fault types are a set of types that combine common faults in the power system, which may include single-phase ground faults or transformer inter-turn short circuits. Fault injection can be completed through the fault simulation module built into the model. During the simulation, a simulation step size of 1ms or 2ms can be selected. The full-time simulation data sequence is a complete data chain covering the steady state before the fault occurs, the fault occurrence process, the fault duration stage, and the recovery stage after the fault is cleared.

[0024] Multiple time-series feature vectors are extracted from the full-time simulation data sequence. The extraction process can first use a sliding window segmentation process, with the window length selectable as 0.5 seconds or 1 second, to divide the full-time data into multiple continuous subsequences. Then, feature information that reflects the time-series change pattern of the data is extracted from each subsequence and integrated to form a time-series feature vector.

[0025] The matching degree between the real-time running data sequence and the full-time simulation data sequence for each preset fault type is calculated on multiple time-series feature vectors to obtain multiple matching degree indices. The calculation process involves comparing the degree of agreement between the corresponding time-series feature vectors of the two sets of data, obtaining quantified matching values ​​through a preset calculation method, and then mapping these values ​​to values ​​between 0 and 1 using a normalization function. The normalization function can be expressed as y=(x max -x) / (x max -x_ min ), where x is the original matching value, x max and x min These are the maximum and minimum values ​​of all matching values ​​under this feature vector, respectively.

[0026] Multiple matching indicators are fused according to preset weights to generate a comprehensive similarity evaluation value. The preset weights can be set according to the importance of different feature vectors to fault diagnosis. For example, the weight of some key feature vectors can be set to 0.2 or 0.3. The fusion process is to multiply each matching indicator by its corresponding weight and then sum them to obtain the final comprehensive similarity evaluation value.

[0027] The comprehensive similarity evaluation value is input into the pre-trained fault classification model, which outputs the actual fault type diagnosis result of the power system. The pre-trained fault classification model is a model trained with a large amount of historical fault data, such as a BP neural network. It can determine the fault type based on the input comprehensive similarity evaluation value. After input, the model will output the probability of each preset fault type, select the one with the highest probability as the candidate result, and check whether the probability exceeds the preset confidence threshold. The confidence threshold can be selected as 0.7 or 0.8.

[0028] In existing technologies, fault diagnosis often uses simulation models with fixed parameters, which cannot be adjusted according to the real-time operating status of the system. Only a small number of common fault types can be injected, feature extraction is mostly limited to local data after the fault occurs, matching degree evaluation mostly uses a single index, and the classification model has weak generalization ability.

[0029] Compared with existing technologies, the method in this embodiment dynamically calibrates the digital twin model through an adaptive parameter estimation algorithm, enabling the model to adapt to the real-time operating state of the power system and avoiding simulation deviations caused by fixed parameter models. By injecting multiple preset fault types and generating a full-time simulation data sequence, combined with the extraction of time-series feature vectors, the complete evolution process of the fault is comprehensively captured, overcoming the limitations of local data diagnosis. Through the weighted fusion of multiple matching degree indicators and the application of a pre-trained classification model, the accuracy and reliability of fault diagnosis are improved. These processing methods work together to solve the technical problems of poor model adaptability, incomplete feature extraction, insufficient diagnostic accuracy, and weak generalization ability in existing technologies, thus improving the overall effect of fault diagnosis.

[0030] Preferably, in S1, voltage data sequences, current data sequences, and frequency data sequences from multiple monitoring points are synchronously collected within a preset time window to form an original operating data set; the original operating data set is subjected to real-time quality assessment and screening, including data integrity verification, validity verification, and consistency verification; for abnormal data points that fail the verification, they are filled and corrected based on historical data from the same monitoring point or real-time data from adjacent monitoring points to generate a real-time operating data sequence.

[0031] For example, when collecting real-time operational data sequences, voltage, current, and frequency data sequences from multiple monitoring points within a preset time window are simultaneously collected to form the original operational data set. These multiple monitoring points can be located at power plant outgoing lines, substation busbars, and important load connection points; the number can be selected as 10 or 15. Synchronous acquisition can be achieved through GPS timing, ensuring that the timestamp error of the data from each monitoring point is within 1ms. The original operational data set is then subjected to real-time quality assessment and filtering. Data integrity verification in the quality assessment checks the proportion of missing data at each monitoring point within the preset time window; for example, a missing proportion not exceeding 5% is considered a pass. Validity verification determines whether the data is within a reasonable range; for example, the valid range for voltage data can be 85% to 115% of the rated voltage. Consistency verification compares the deviation of similar data from adjacent monitoring points at the same time; a deviation not exceeding 3% is considered a pass. For abnormal data points that fail verification, if it is a single missing data point from a single monitoring point, it can be filled by interpolation based on historical data from the same period within the past 3 months for that monitoring point. Linear interpolation can be used as the interpolation method. If it is a continuous anomaly caused by a monitoring point failure, it can be corrected by using the real-time data of two adjacent monitoring points through a weighted average method. The weights can be allocated according to the distance between the monitoring point and the failure point. The weights for points that are close can be 0.6 or 0.7. Finally, a complete and reliable real-time operation data sequence is generated.

[0032] This embodiment ensures the spatial correlation of data through synchronous acquisition from multiple monitoring points, comprehensively guarantees the validity of data through triple quality verification, and avoids data loss through targeted abnormal data filling and correction. It solves the problems of incomplete data acquisition, lack of quality assurance, and rough anomaly handling in the prior art, and provides a reliable data foundation for the accuracy of subsequent diagnosis. This refined processing of the entire data process reflects the improvement of technology.

[0033] Preferably, S2 includes the following sub-steps: S21: Input the real-time running data sequence into the pre-established digital twin model of the power system, and dynamically identify and adjust the equivalent parameters of key components in the digital twin model of the power system based on the online recursive parameter identification algorithm. Key components include generators, lines and loads. S22: Run the digital twin model of the power system using the adjusted parameters to generate the corresponding simulation output sequence; S23: Calculate the normalized root mean square error between the simulation output sequence and the real-time running data sequence within the same time period. If the error is greater than the preset convergence threshold, repeat S21 and S22 for iterative calibration. S24: When the error is less than or equal to the convergence threshold, the model under the current parameters will be locked as the calibrated dynamic simulation model.

[0034] For example, the process of adjusting the parameters of a digital twin model to generate a dynamic simulation model includes several steps. First, the real-time running data sequence is input into a pre-established power system digital twin model, which includes the equivalent parameters of key components such as generators, lines, and loads. Then, based on an online recursive parameter identification algorithm, the equivalent parameters of these key components are dynamically identified and adjusted. The online recursive parameter identification algorithm can be a recursive least squares method with a forgetting factor. The equivalent parameters of generators can include transient reactance and subtransient reactance; the equivalent parameters of lines can include resistance and reactance; and the equivalent parameters of loads can include active power and reactive power. Next, the power system digital twin model is run using the adjusted parameters to generate a corresponding simulation output sequence. The simulation output sequence maintains the same time period as the real-time running data sequence, both being preset 5 minutes or 10 minutes. Then, the normalized root mean square error (RMSE) of both is calculated. The calculation process is as follows: first, the sum of squares of the differences between the simulated value and the real-time value at each time step is calculated; then, this sum is divided by the sum of squares of the real-time values ​​to obtain the RMSE; finally, this sum is divided by the maximum real-time value to obtain the normalized RMSE. A preset convergence threshold of 0.05 can be selected. If the calculated error is greater than this threshold, the parameter identification and model running steps are repeated for iterative calibration. When the error is less than or equal to the convergence threshold, the model under the current parameters is locked as the calibrated dynamic simulation model. The number of iterations usually does not exceed 10 to meet the convergence requirement.

[0035] In existing technologies, digital twin models often use fixed parameters or only perform offline parameter calibration without iterative convergence judgment, resulting in significant deviations between the model and the actual operating state. Specifically, they may fix the generator transient reactance at 0.2 pu and the line resistance at 0.01 Ω / km, without adjusting based on real-time data, performing parameter settings only once during model establishment, and not calculating the error between simulation and real-time data. Compared with existing technologies, the method in this embodiment dynamically identifies key component parameters through an online recursive algorithm, and ensures model accuracy through iterative calibration and convergence judgment. This solves the problem of fixed model parameters and inability to adapt to the dynamic operating state of the system in existing technologies, enabling the simulation model to accurately match the actual system and providing a reliable model foundation for subsequent fault simulation.

[0036] More preferably, in S21, a recursive least squares method with a forgetting factor is used as the online recursive parameter identification algorithm; a sliding time window of preset length is used to extract the real-time running data sequence to construct the parameter identification input-output data pair of key components; for each key component to be identified, the parameter variability of the key component is calculated based on the deviation between its historical running data and simulation data; based on the parameter variability, the subset of key components that need parameter updates within the current time window is dynamically determined; recursive parameter identification and adjustment are performed only on the components in the subset of key components to update their equivalent parameters in the power system digital twin model.

[0037] For example, when dynamically identifying key component parameters, a recursive least squares method with a forgetting factor is used as the online recursive parameter identification algorithm. The forgetting factor can be selected as 0.95. This algorithm achieves dynamic parameter updates by assigning higher weights to new data and lower weights to old data. Then, a sliding time window of a preset length is used to extract the real-time running data sequence. The sliding time window length can be 1 minute or 2 minutes, and each sliding step can be 30 seconds. The data extracted in this window is used to construct the input-output data pairs for key component parameter identification. The input data can be voltage or current data, and the output data can be power data. For each key component to be identified, the parameter variability of the key component is calculated based on the deviation between its historical running data and simulation data. The calculation process involves finding the standard deviation of the deviation within the last 10 sliding windows; a larger standard deviation indicates a higher parameter variability. A threshold is set based on the parameter variability, for example, the threshold can be 0.03 or 0.05. The key components with variability greater than the threshold are grouped into a subset of key components that need parameter updates. Recursive parameter identification and adjustment are performed only on the components in the subset. For example, only the line reactance or load power parameters with excessive variability are updated, and their equivalent parameters in the power system digital twin model are updated. The parameters of components not included in the subset remain unchanged.

[0038] In existing technologies, parameter identification often involves uniformly updating all key components without distinguishing parameter variability, resulting in high computational load and low calibration efficiency. Specifically, this method can synchronously and recursively identify the parameters of all components in the generator, line, and load, without setting a sliding time window, directly calculating based on all historical data, and disregarding differences in parameter variability. Compared with existing technologies, the method in this embodiment improves the timeliness of parameter updates through an algorithm with a forgetting factor, extracts effective data through a sliding time window, and filters the subset of components requiring updates based on parameter variability. This solves the problems of high computational load, low efficiency, and poor timeliness in existing parameter identification technologies, reducing computational costs while ensuring model accuracy.

[0039] Preferably, in S3, multiple preset fault types to be injected are configured on the calibrated dynamic simulation model, and the fault parameters, injection time, and duration of each fault type are set; the simulation environment of the calibrated dynamic simulation model is initialized according to the running status of the real-time running data sequence at the end of the preset time window; the injection and simulation of each preset fault type are executed sequentially with a set simulation step size, and the simulation data of the entire process from the pre-fault steady state, fault occurrence, fault duration to fault clearance and recovery are recorded synchronously; the full-process simulation data corresponding to each preset fault type are integrated and time-aligned to generate a full-time simulation data sequence.

[0040] For example, on the calibrated dynamic simulation model, multiple preset fault types to be injected are first configured. These preset fault types include single-phase ground faults and transformer inter-turn short circuits, as well as two-phase short circuits and generator demagnetization faults. Fault parameters for each fault type need to be set individually. For instance, for a two-phase short circuit, the fault phases can be selected as A and B, the transition resistance can be selected as 0.5Ω or 1Ω, the injection time can be selected as 1 or 2 seconds after the end of the preset time window, and the duration can be selected as 0.1 or 0.2 seconds. Then, based on the real-time running data sequence at the end of the preset time window, the simulation environment of the calibrated dynamic simulation model is initialized. Initialization parameters include the voltage amplitude and phase of each node, the current of each branch, and the system frequency, ensuring that the initial simulation state is consistent with the actual system operating state. During simulation, a simulation step size of 1ms or 2ms can be selected to sequentially inject and simulate each preset fault type. Simultaneously, simulation data is recorded throughout the entire process, from the pre-fault steady state, fault occurrence, fault duration, to fault clearing and recovery. The fault clearing method can be circuit breaker tripping, and the recovery process involves the system naturally transitioning to a new steady state after fault clearing. Finally, the full-process simulation data corresponding to each preset fault type are integrated and time-aligned based on timestamps to ensure the temporal consistency of data at different stages, generating a full-time-series simulation data sequence.

[0041] In existing technologies, fault injection often involves randomly selecting the injection time, without initializing the simulation environment, and only recording local data after the fault occurs, lacking full-time data integration. Specifically, this method can randomly select any time to inject the fault, without considering the actual system operating state, using a simulation step size of 5ms, recording only data within 0.5 seconds after the fault occurs, and without time alignment. Compared with existing technologies, the method in this embodiment, by precisely configuring fault parameters and injection time, initializing the simulation environment based on the actual system state, recording full-cycle fault data, and performing time alignment, solves the problems of fault simulation being disconnected from the actual scenario, incomplete data, and disordered timing in existing technologies. It provides comprehensive and orderly simulation data for subsequent feature extraction, and this full-process fault simulation design reflects a technological improvement.

[0042] Preferably, in S4, the full-time simulation data sequence is segmented by a sliding window to obtain multiple simulation data subsequences; for each simulation data subsequence, multiple initial feature indices in the time domain, frequency domain, and time-frequency domain are calculated respectively; the same initial feature indices of all simulation data subsequences under the same fault type are arranged in chronological order to form the time-series variation curve of the feature index; from all time-series variation curves, several curves with the highest correlation with typical fault modes are selected, and their corresponding feature indices are used as basic features; based on the basic features, dimensionality reduction and reconstruction are performed by principal component analysis to generate multiple time-series feature vectors corresponding to each preset fault type.

[0043] For example, when extracting multiple time-series feature vectors from a full-time simulation data sequence, the full-time simulation data sequence is first segmented by a sliding window. The sliding window length can be selected as 0.5 seconds, and the sliding step size can be selected as 0.25 seconds, resulting in multiple non-overlapping or partially overlapping simulation data subsequences. For each simulation data subsequence, multiple initial feature indices are calculated in the time domain, frequency domain, and time-frequency domain. The time-domain indices can include peak value, mean, variance, and kurtosis. The calculation process is to find the maximum value of all data points in each subsequence to obtain the peak value, find the arithmetic mean to obtain the mean, and find the average of the squares of the deviations of the data points from the mean to obtain the variance. The frequency-domain indices can be obtained through Fast Fourier Transform (FFT). First, the data subsequence is padded with zeros to a length that is an integer power of 2, and then FFT is performed to extract indices such as fundamental frequency and harmonic content. The time-frequency domain indices can be obtained through wavelet transform. The db4 wavelet is selected as the mother wavelet, and the number of decomposition layers can be selected as 3 or 4 layers. The energy values ​​of the detail coefficients of each layer are extracted as indices. Then, the same initial characteristic indices of all simulation data subsequences under the same fault type are arranged in chronological order to form the time-series variation curves of these characteristic indices. By calculating the Pearson correlation coefficient between each curve and the typical fault mode curve, several curves with a correlation coefficient greater than 0.8 are selected, and their corresponding characteristic indices are used as basic features. Based on the basic features, dimensionality reduction and reconstruction are performed using principal component analysis. The covariance matrix of the basic features is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The top few eigenvectors with a cumulative contribution rate of 85% are selected as principal components, thereby generating multiple time-series feature vectors corresponding to each preset fault type.

[0044] In existing technologies, feature extraction often focuses on single-domain indicators, lacks sliding window segmentation, and suffers from a lack of basic feature selection, resulting in simplistic and crude dimensionality reduction methods. Specifically, it may only calculate the time-domain peak and mean values ​​of fault data, without performing frequency-domain and time-frequency domain analysis, without constructing time-series variation curves, and directly use random forests for feature selection, leading to poor dimensionality reduction performance. Compared with existing technologies, the method in this embodiment solves the problems of incomplete feature extraction, weak time-series correlation, and poor dimensionality reduction performance in existing technologies by using multi-domain initial feature extraction, sliding window segmentation to construct time-series curves, correlation analysis to select basic features, and principal component analysis for precise dimensionality reduction. The resulting time-series feature vector can accurately characterize the fault evolution pattern, demonstrating a technological improvement.

[0045] Preferably, in S5, for each preset fault type, multiple time-series feature vectors corresponding to the real-time running data sequence and the full-time simulation data sequence are extracted respectively; for each time-series feature vector, a dynamic time warping algorithm is used to calculate the minimum warping distance between the feature vector time sequence of the real-time running data sequence and the corresponding feature vector time sequence of the full-time simulation data sequence; the minimum warping distance is mapped to a value between 0 and 1 through a preset normalization function, which is used as the single matching degree under the time-series feature vector; for each preset fault type, its single matching degree on all time-series feature vectors is summarized to form multiple matching degree indices for the fault type.

[0046] For example, when calculating the matching degree between the real-time running data sequence and the simulation data of each preset fault type, firstly, for each preset fault type, multiple corresponding time-series feature vectors are extracted from the real-time running data sequence and the full-time simulation data sequence of that type, ensuring that the feature vector dimensions and types of both are consistent. For each time-series feature vector, a dynamic time warping algorithm is used to calculate the matching degree. This algorithm first constructs a distance matrix between the real-time feature vector time series and the simulation feature vector time series, where Euclidean distance can be used. Then, dynamic programming is used to find the optimal matching path from the upper left corner to the lower right corner of the distance matrix. The path constraint can be that the maximum warp length is 20% or 30% of the sequence length, and the sum of the distances on the path is the minimum warping distance. Then, the minimum warping distance is mapped to a value between 0 and 1 using a preset normalization function. For each preset fault type, its single matching degree on all time-series feature vectors is summarized to constitute multiple matching degree indices for that fault type.

[0047] In existing technologies, matching degree calculation often uses direct comparison with Euclidean distance, without considering the time offset of time-series data, making it difficult to distinguish similar faults. Compared with existing technologies, the method in this embodiment solves the time offset problem of time-series data through a dynamic time warping algorithm, which solves the problems of inaccurate matching degree calculation and weak ability to distinguish similar faults in existing technologies, and improves the reliability of matching degree evaluation.

[0048] Preferably, in S6, multiple matching degree indices corresponding to each preset fault type are obtained; a matching degree index matrix is ​​constructed, wherein rows correspond to different preset fault types, columns correspond to different time-series feature vectors, and matrix elements are the corresponding matching degree indices; for each time-series feature vector, the discrimination and dispersion of its corresponding matching degree index among all preset fault types are calculated; based on the entropy weight method, the fusion weight of each time-series feature vector is dynamically calculated according to the discrimination and dispersion, and used as the preset weight; using the calculated fusion weight, the matching degree indices of each row in the matching degree index matrix are weighted and summed to generate a comprehensive similarity evaluation value corresponding to each preset fault type.

[0049] For example, when fusing multiple matching indices to generate a comprehensive similarity evaluation value, firstly, multiple matching indices corresponding to each preset fault type are obtained. Assuming there are 8 preset fault types and 6 temporal feature vectors, an 8x6 matching index matrix is ​​constructed. The element in the i-th row and j-th column of the matrix represents the matching index of the i-th fault type under the j-th temporal feature vector. For each temporal feature vector, the discriminancy and dispersion of its corresponding matching index among all preset fault types are calculated. The discriminancy can be obtained by calculating the coefficient of variation of each index; a larger coefficient of variation indicates higher discriminancy. The dispersion can be obtained by calculating the interquartile range of the index. Based on the entropy weight method, the fusion weight of each temporal feature vector is dynamically calculated according to the discriminancy and dispersion. The calculation process involves first normalizing the matching index of each temporal feature vector to obtain a normalized matrix p. ij , where p ij =x ij / Σx ij (x) ij (For the matching degree of the j-th feature of the i-th type of fault), then calculate the information entropy H of each feature. j =-kΣp ij lnp ij (k=1 / lnm, m is the number of fault types), and finally calculate the weight w. j =(1-H j ) / Σ(1-H j This weight is the preset fusion weight. Using the calculated fusion weight, the matching degree indicators in each row of the matching degree indicator matrix are weighted and summed, i.e., the comprehensive similarity evaluation value = Σ(w j ×x_ ij This generates a comprehensive similarity evaluation value for each preset fault type.

[0050] In existing technologies, matching degree fusion often employs an equal-weighted summation method, neglecting to calculate feature discriminativeness and dispersion, and the weighting is subjective. Specifically, it assigns equal weights to the matching degree indices of all time-series feature vectors and directly sums them to obtain the comprehensive similarity, without considering the differences in the contribution of different features to fault differentiation. Compared with existing technologies, the method in this embodiment constructs a matching degree index matrix to organize data relationships and dynamically calculates the fusion weights based on discriminativeness and dispersion using the entropy weight method. This solves the problems of subjective weighting and inaccurate fusion results in existing technologies, improving the scientific rigor of the comprehensive similarity assessment.

[0051] Preferably, in S7, a pre-trained fault classification model based on a deep neural network is constructed, with the number of input layer nodes being the same as the dimension of the comprehensive similarity evaluation value, and the number of output layer nodes being the same as the number of preset fault types. Simulation data from historical multi-fault scenarios is used to generate a sample set of comprehensive similarity evaluation values ​​for different fault types and their corresponding real fault type labels, which are then used to train and cross-validate the fault classification model. The comprehensive similarity evaluation values ​​are input into the pre-trained fault classification model, and its softmax layer outputs the probability distribution corresponding to each preset fault type. The fault type with the highest probability value is selected as the candidate diagnostic result, and it is checked whether this probability value exceeds a preset confidence threshold. If it exceeds the threshold, the candidate diagnostic result is output as the actual fault type diagnostic result of the power system; if it does not exceed the threshold, an indication signal of "unknown or compound fault" is output.

[0052] For example, when outputting fault type diagnosis results, a pre-trained fault classification model based on a deep neural network is first constructed. The model can be a CNN-LSTM hybrid model, with the number of input layer nodes matching the dimension of the comprehensive similarity evaluation value. If the comprehensive similarity evaluation value is 6-dimensional, then the number of input layer nodes is 6. The number of output layer nodes is the same as the number of preset fault types; if there are 8 preset fault types, then the number of output layer nodes is 8. During model training, simulation data from historical multi-fault scenarios is used to generate a sample set of comprehensive similarity evaluation values ​​for different fault types. The number of sample sets can be 1000 or 2000 sets. Each sample set is labeled with the corresponding real fault type label. A 5-fold or 10-fold cross-validation method is used to train the fault classification model to avoid overfitting. After inputting the comprehensive similarity evaluation value into the pre-trained fault classification model, the model outputs the probability distribution corresponding to each preset fault type through a softmax layer, where the sum of all elements in the probability distribution is 1. The fault type with the highest probability value is selected as the candidate diagnostic result. At the same time, it is checked whether the probability value exceeds the preset confidence threshold. The confidence threshold can be selected as 0.7 or 0.8. If it exceeds the threshold, the candidate diagnostic result is output as the actual fault type diagnostic result of the power system. If it does not exceed the threshold, an indication signal of "unknown or compound fault" is output, indicating that further investigation is required.

[0053] Compared with existing technologies, the method in this embodiment improves classification accuracy through deep neural networks, ensures model generalization ability through a large number of samples and cross-validation, and avoids misjudgment through confidence threshold judgment. It solves the problems of low classification accuracy, weak generalization ability and insufficient identification of unknown faults in existing technologies, and improves the reliability and robustness of fault diagnosis.

[0054] Example 2 This embodiment is based on embodiment 1: This embodiment provides an intelligent simulation diagnosis system for power faults, including: The dynamic simulation model building module is configured to collect real-time operating data sequences of the power system within a preset time window, combine the real-time operating data sequences with an adaptive parameter estimation algorithm, adjust the parameters of the pre-established digital twin model of the power system, and generate a calibrated dynamic simulation model. The time-series feature vector extraction module is configured to sequentially inject multiple preset fault types into the dynamic simulation model, run the simulation to generate a full-time simulation data sequence corresponding to each preset fault type, and then extract multiple time-series feature vectors from the full-time simulation data sequence. The matching degree index calculation module is configured to calculate the matching degree between the real-time running data sequence and the full-time simulation data sequence of each preset fault type on multiple time-series feature vectors, and obtain multiple matching degree indices. The fault type diagnosis result output module is configured to fuse multiple matching degree indicators according to preset weights to generate a comprehensive similarity evaluation value, and input the comprehensive similarity evaluation value into a pre-trained fault classification model to output the actual fault type diagnosis result of the power system.

[0055] Example 3 This embodiment is based on embodiment 1: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the intelligent simulation diagnosis method for power faults described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0056] Example 4 This embodiment is based on embodiment 1: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the intelligent simulation diagnosis method for power faults described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0057] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0058] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for intelligent simulation diagnosis of power faults, characterized in that, include: Collect real-time operating data sequences of the power system within a preset time window, combine the real-time operating data sequences with an adaptive parameter estimation algorithm, adjust the parameters of the pre-established digital twin model of the power system, and generate a calibrated dynamic simulation model. In the dynamic simulation model, multiple preset fault types are injected sequentially, and the simulation is run to generate a full-time simulation data sequence corresponding to each preset fault type. Then, multiple time-series feature vectors are extracted from the full-time simulation data sequence. The matching degree between the real-time running data sequence and the full-time simulation data sequence for each preset fault type is calculated on multiple time-series feature vectors to obtain multiple matching degree indices; Multiple matching indicators are fused according to preset weights to generate a comprehensive similarity evaluation value. The comprehensive similarity evaluation value is then input into a pre-trained fault classification model to output the actual fault type diagnosis result of the power system.

2. The intelligent simulation diagnosis method for power faults according to claim 1, characterized in that, The data sequence of real-time operation of the power system within a preset time window includes: Synchronously collect real-time operational data sequences from multiple monitoring points within a preset time window to form an original operational data set. The real-time operational data sequences include voltage data sequences, current data sequences, and frequency data sequences. Real-time quality assessment and filtering are performed on the original operational data set, including data integrity verification, validity verification, and consistency verification. For abnormal data points that fail the verification, they are filled and corrected based on historical data from the same monitoring point or real-time data from adjacent monitoring points to generate a real-time running data sequence.

3. The intelligent simulation diagnosis method for power faults according to claim 1, characterized in that, The process of combining real-time operational data sequences and adaptive parameter estimation algorithms to adjust the parameters of a pre-established digital twin model of the power system and generate a calibrated dynamic simulation model includes: The real-time running data sequence is input into the pre-established digital twin model of the power system, and the equivalent parameters of key components in the digital twin model of the power system are dynamically identified and adjusted based on the online recursive parameter identification algorithm. The key components include generators, lines and loads. Run the digital twin model of the power system using the adjusted parameters to generate the corresponding simulation output sequence; Calculate the normalized root mean square error between the simulation output sequence and the real-time running data sequence within the same time period. If the error is greater than the preset convergence threshold, repeat the above two steps for iterative calibration. When the error is less than or equal to the convergence threshold, the model under the current parameters is locked as the calibrated dynamic simulation model.

4. The intelligent simulation diagnosis method for power faults according to claim 3, characterized in that, The method of dynamically identifying and adjusting the equivalent parameters of key components in the digital twin model of a power system based on an online recursive parameter identification algorithm includes: The recursive least squares method with a forgetting factor is used as the online recursive parameter identification algorithm; The real-time running data sequence is captured by a sliding time window of preset length, and the parameter identification input-output data pairs of key components are constructed. For each key component to be identified, the parameter variability of the key component is calculated based on the deviation between its historical operating data and simulation data; Based on the parameter variability, dynamically determine the subset of key components that need parameter updates within the current time window; Recursive parameter identification and adjustment are performed only on components in a subset of key components to update the equivalent parameters in the power system digital twin model.

5. The intelligent simulation diagnosis method for power faults according to claim 1, characterized in that, The process of sequentially injecting multiple preset fault types into the dynamic simulation model and running the simulation to generate a full-time simulation data sequence corresponding to each preset fault type includes: On the calibrated dynamic simulation model, configure multiple preset fault types to be injected, and set the fault parameters, injection time and duration for each fault type; The simulation environment of the calibrated dynamic simulation model is initialized based on the running status of the real-time running data sequence at the end of the preset time window. With the set simulation step size, the injection and simulation of each preset fault type are executed sequentially, and the simulation data of the entire process from the pre-fault steady state, the fault occurrence, the fault duration to the fault clearance and recovery are recorded synchronously. The full-process simulation data corresponding to each preset fault type is integrated and time-aligned to generate a full-time simulation data sequence.

6. The intelligent simulation diagnosis method for power faults according to claim 1, characterized in that, The extraction of multiple time-series feature vectors from the full-time simulation data sequence includes: The full-time simulation data sequence is segmented by a sliding window to obtain multiple simulation data subsequences; For each subsequence of simulation data, calculate multiple initial feature indices in the time domain, frequency domain, and time-frequency domain respectively; Arrange the same initial characteristic index of all simulation data subsequences under the same fault type in chronological order to form the time series change curve of the characteristic index; From all the time-series variation curves, select the curves with the highest correlation to typical failure modes, and use their corresponding characteristic indicators as the basic features. Based on the aforementioned basic features, dimensionality reduction and reconstruction are performed using principal component analysis to generate multiple time-series feature vectors corresponding to each preset fault type.

7. The intelligent simulation diagnosis method for power faults according to claim 1, characterized in that, The process involves calculating the matching degree between the real-time running data sequence and the full-time simulation data sequence for each preset fault type on multiple time-series feature vectors, resulting in multiple matching degree indices, including: For each preset fault type, multiple time-series feature vectors corresponding to the real-time running data sequence and the full-time simulation data sequence are extracted respectively; For each time-series feature vector, a dynamic time warping algorithm is used to calculate the minimum warping distance between the feature vector time series of the real-time running data sequence and the corresponding feature vector time series of the full-time simulation data sequence. The minimum normalization distance is mapped to a value between 0 and 1 using a preset normalization function, and is used as the single matching degree under this time series feature vector; For each preset fault type, its single matching degree on all time-series feature vectors is summarized to form multiple matching degree indices for that fault type.

8. The intelligent simulation diagnosis method for power faults according to claim 1, characterized in that, The step of fusing multiple matching metrics according to preset weights to generate a comprehensive similarity evaluation value includes: Obtain multiple matching metrics corresponding to each preset fault type; Construct a matching degree index matrix, where rows correspond to different preset fault types, columns correspond to different time-series feature vectors, and matrix elements are the corresponding matching degree indices; For each time-series feature vector, calculate its corresponding matching degree index, which has the discriminant and discreteness among all preset fault types; Based on the entropy weight method, the fusion weight of each time series feature vector is dynamically calculated according to the discrimination and dispersion, and used as the preset weight; Using the calculated fusion weights, the matching degree indices in each row of the matching degree index matrix are weighted and summed to generate a comprehensive similarity evaluation value for each preset fault type.

9. The intelligent simulation diagnosis method for power faults according to claim 1, characterized in that, The process of inputting the comprehensive similarity evaluation value into the pre-trained fault classification model and outputting the actual fault type diagnosis result of the power system includes: Construct a pre-trained fault classification model based on a deep neural network, where the number of nodes in the input layer is the same as the dimension of the comprehensive similarity evaluation value, and the number of nodes in the output layer is the same as the number of preset fault types. Using simulation data from historical multi-fault scenarios, a comprehensive similarity evaluation value sample set for different fault types and their corresponding real fault type labels are generated to train and cross-validate the fault classification model. The comprehensive similarity evaluation value is input into the pre-trained fault classification model, and its softmax layer outputs the probability distribution corresponding to each preset fault type. The fault type with the highest probability value is selected as the candidate diagnostic result, and it is checked whether the probability value exceeds the preset confidence threshold. If it exceeds the threshold, the candidate diagnostic result is output as the actual fault type diagnostic result of the power system. If it does not exceed the threshold, an indication signal for an unknown or compound fault is output.

10. A power fault intelligent simulation diagnosis system, characterized in that, include: The dynamic simulation model building module is configured to collect real-time operating data sequences of the power system within a preset time window, combine the real-time operating data sequences with an adaptive parameter estimation algorithm, adjust the parameters of the pre-established digital twin model of the power system, and generate a calibrated dynamic simulation model. The time-series feature vector extraction module is configured to sequentially inject multiple preset fault types into the dynamic simulation model, run the simulation to generate a full-time simulation data sequence corresponding to each preset fault type, and then extract multiple time-series feature vectors from the full-time simulation data sequence. The matching degree index calculation module is configured to calculate the matching degree between the real-time running data sequence and the full-time simulation data sequence of each preset fault type on multiple time-series feature vectors, and obtain multiple matching degree indices. The fault type diagnosis result output module is configured to fuse multiple matching degree indicators according to preset weights to generate a comprehensive similarity evaluation value, and input the comprehensive similarity evaluation value into a pre-trained fault classification model to output the actual fault type diagnosis result of the power system.

Images